betaFeedback

WorksPhysics and Astronomy

Enabling joint chemical and meteorological data assimilation with an atmospheric foundation model

Abstract

The hydroxyl radical (OH) is the main oxidant in the troposphere and accounts for 88% of the methane sink, yet its abundance and trends remain poorly constrained. Satellites observe many of the chemical and meteorological drivers of OH, but assimilating both requires a model that represents chemistry and meteorology together. Existing chemistry-climate models are too expensive for joint data assimilation. Here, we present AuroraChem, a fine-tuned version of the Aurora atmospheric foundation model that jointly simulates meteorology and the GEOS-Chem chemical fields needed to constrain OH. A one-year AuroraChem rollout takes 12 minutes on one GPU. Free rollouts remain stable for one year but do not fully capture seasonal cycles. Using idealized data assimilation, we find that chemical observations can constrain the meteorological state and meteorological observations can constrain the chemical state. Inserting only the chemical fields raises the annual mean spatial correlation of 500 hPa temperature from 0.89 to 0.92 and recovers the moisture and circulation pattern of the Indian summer monsoon. Inserting only the meteorological fields raises the 500 hPa OH correlation from 0.89 to 0.93, close to the 0.94 from inserting the chemical fields themselves, and recovers the OH seasonal cycle. Inserting only NO2 columns sampled at TROPOMI pixels gives an OH correlation of 0.93. These relationships likely reflect learned covariances rather than physical coupling, and covariances are what data assimilation uses to let an observation of one variable update the others. AuroraChem therefore provides a new pathway to joint chemical and meteorological data assimilation.

Introduction

The hydroxyl radical (OH) is the primary oxidant in the troposphere, oxidizing methane, carbon monoxide, and VOCs to produce ozone and other reactive species. It governs the lifetime of tropospheric pollutants and ozone-depleting substances while also exerting an indirect control on the long-term radiative budget through its regulation of the atmospheric lifetime of gases such as methane (e.g., [15]). As such, OH is arguably the single most important reactive molecule in the troposphere.

Despite its importance, there are large uncertainties in both OH abundance and long-term trends. Global mean OH across 14 chemistry-climate models ranges from 8.7 to 12.8×10512.8 \times10^{5} molec cm−3^{-3} from 2000–2010 ([39]). Direct OH observations are spatially sparse and have uncertainties of ±30%\pm30\% (e.g., [31]). In lieu of direct observations, observations of methyl chloroform (CH3_3CCl3_3) have long been used as the main constraint on OH abundance (e.g., Montzka et al., 2011; Rigby et al., [27]; Turner et al., [32]) and, as such, the oxidative capacity of the troposphere. Methyl chloroform concentrations have declined to near zero following the Montreal Protocol, which weakens this constraint [16]. As a result, observational constraints on OH are weaker today than they were 20 years ago.

Satellites observe many of the physical and chemical drivers of OH. Murray et al. [24, 25] summarize the drivers of global mean tropospheric OH as

[OH]∝JO3qSNSC3/2.(1)[\mathrm{OH}] \propto J_{\mathrm{O}_3}q\frac{S_N}{S_C^{3/2}} . \tag*{(1)}

where JO3J_{\mathrm{O}_3} is the photolysis frequency of ozone, qq is specific humidity, and SNS_N and SCS_C are the sources of reactive nitrogen and reactive carbon, respectively. Ozone photolysis in the presence of water vapor is the primary source of OH. Reactive nitrogen increases OH by recycling peroxy radicals back to OH, and reactive carbon consumes OH. Drivers of global mean OH co-vary with physical meteorology and climate, in which Turner et al. [33] and Mei et al. [19] showed that OH variability can be explained by the ENSO index. Most of these drivers are well-observed by existing satellites with daily global coverage and have potential to be used for reconstructing spatially-resolved OH fields. For example, satellites observe water vapor (e.g., AIRS), NO2\mathrm{NO}_2 (e.g., TROPOMI), ozone (e.g., OMI), and CO (e.g., MOPITT) with near-daily global coverage. Prior work has used these observations to estimate OH. Wolfe et al. [38] estimated total column OH from satellite HCHO retrievals with a steady-state approximation. Anderson et al. [1] trained a neural network to predict total column OH from chemistry-climate model fields (including NO2\mathrm{NO}_2, CO, HCHO, H2O\mathrm{H}_2\mathrm{O}, and O3\mathrm{O}_3), then applied it to satellite observations of the same species. However, these satellite-based estimates are limited by the spatiotemporal sparsity of satellite retrievals. Total column OH from Wolfe et al. [38] is limited to open oceans where satellite retrievals of HCHO are reliable and available at monthly time resolution due to temporal averaging. Similarly, the total column OH estimates from Anderson et al. [1] are only available over tropical ocean areas with monthly temporal resolution.

Data assimilation (DA) can combine these observations with a model to produce continuous, global OH fields. Miyazaki et al. [20] assimilated satellite atmospheric composition observations into an ensemble of chemical transport models (CTMs) and reduced the inter-model spread in regional mean OH by 16–50%. However, CTMs like those used in Miyazaki et al. [20] are driven by offline meteorological inputs, so their chemical state does not feedback to their climate state. Joint assimilation requires a model that represents chemistry and meteorology in a single state, along with the covariances between them. Coupled chemistry-climate models (CCMs) provide this but are computationally expensive. As an example, adding interactive chemistry to CESM2, a state-of-the-art climate model, increases its computational cost by a factor of seven [7]. Ensemble DA multiplies this cost by the number of ensemble members, typically 10–100 [10]. Prior efforts to assimilate chemistry to constrain meteorology have been limited in scope. Saide et al. [28] assimilated satellite atmospheric composition observations to constrain radiative fields that control boundary layer stability, and Liu et al. [17] assimilated synthetic satellite NO2\mathrm{NO}_2 observations to estimate boundary layer winds. However, both studies conducted assimilation on regional domains and covered a time range of days. Joint chemical and meteorological DA with a CCM is therefore computationally prohibitive at global scale and over the multi-year time periods necessary to constrain OH trends.

Machine learning (ML) emulators could make joint DA affordable. ML weather prediction models such as GraphCast, FourCastNet v3, and AIFS match or outperform the ECMWF Integrated Forecast System (IFS) on medium-range forecasts [14, 5, 21]. ML climate emulators run stably for decades to centuries at orders of magnitude lower cost than numerical climate models [6, 13, 37]. ML emulators of atmospheric chemistry are less mature. Keller and Evans [11] replaced the chemical solver in GEOS-Chem with random forests and achieved stable month-long simulations, but at twice the computational cost of the original solver. Kelp et al. [12] trained a faster neural network emulator of the chemical mechanism, but its simulations became unstable after 1.5 weeks. Mei et al. [19] developed a stable coupled chemistry-climate emulator using a linear inverse model, but it simulates column-average quantities and does not capture the nonlinear chemical and physical processes driving OH. Thus, the emulators described above do not provide a joint chemical and meteorological state, stability over the time-scales necessary for DA, and the chemical species needed to constrain OH.

One emulator that may satisfy the above conditions is Aurora [4]. Aurora is a foundation model pretrained on more than 1 PB of reanalysis, analysis, NWP, and climate model output to learn a general representation of the atmosphere. Fine-tuning adapts the pretrained model to new variables with much smaller training datasets. Bodnar et al. [4] fine-tuned Aurora to forecast air quality (ozone, CO, NOx\mathrm{NO}_{x}, SO2\mathrm{SO}_{2}, and particulate matter) jointly with weather. This is the first ML emulator that jointly simulates chemistry and meteorology on a 3D global grid. However, air quality Aurora does not include OH and the other species needed to constrain oxidative capacity, including Cl and 14CO{}^{14}\mathrm{CO}. It was also evaluated on 5-day forecasts, not on the longer time-scales necessary for DA over seasons.

Here, we fine-tune Aurora with a parsimonious set of chemical and meteorological fields necessary to constrain OH and the atmospheric oxidative capacity. The fine-tuned model, termed “AuroraChem" (http://aurorachem.github.io), is a joint chemistry-meteorology emulator which is accurate for short lead-times and stable over one year. We use an idealized form of DA, “direct insertion," to test whether chemical observations constrain the meteorological state and whether meteorological observations constrain the chemical state. In our tests, periodic meteorological observations can constrain the chemical state and vice versa. We also show that inserting a single chemical species, even when sampled and column-integrated like a TROPOMI retrieval, constrains OH with similar performance to inserting meteorological fields. These results suggest that AuroraChem could serve as the forward model for joint chemical and meteorological DA.

Fine-tuning an atmospheric foundation model with chemistry

AuroraChem architecture

Bodnar et al. [4] pretrained Aurora to learn a general representation of the atmospheric system. The pretraining dataset included fields from 10 reanalyses, analyses, NWP models, and climate models, totaling over 1 petabyte. Bodnar et al. [4] describe the pretraining in detail. The prognostic fields of base Aurora are temperature, specific humidity, zonal and meridional wind, and geopotential height on 13 pressure levels, and 2 m temperature, 10 m zonal and meridional wind, and mean sea level pressure at the surface. Base Aurora also takes static fields as inputs (land-sea mask, soil type, and topography). Aurora ingests the atmospheric states at t0−6 ht_0 - 6\,\mathrm{h} and t0t_0 to forecast the atmospheric state at t0+6 ht_0 + 6\,\mathrm{h} (Figure 1). Longer forecasts, or rollouts, apply the model autoregressively, using each forecast as the input for the next step.

Schematic of AuroraChem architecture

Figure 1. Schematic of AuroraChem architecture. The encoder maps meteorological, chemical, surface, and static fields at t−6 ht - 6\,\mathrm{h} and tt to a latent state with four vertical levels. The backbone advances the latent state by 6 h6\,\mathrm{h}. The decoder maps it back to the output fields at t+6 ht + 6\,\mathrm{h}. Solid arrows show new weights trained from scratch at a high learning rate. Dotted arrows show pretrained weights updated at a low learning rate.

Aurora has three components: an encoder, a backbone, and a decoder (Figure 1). The encoder first divides the atmospheric fields at each pressure level into patches of size P×P×V×TP \times P \times V \times T, where V=5V = 5 is the number of variables and T=2T = 2 is the number of input times. The input patches for each pressure level are embedded into a vector with 512 elements via cross-attention blocks. Crucially, the encoder’s learned embedding layer accepts inputs on different grid resolutions and with different sets of variables, which allows pretraining and fine-tuning on diverse datasets. The 13 atmospheric embedding vectors are then compressed with cross-attention into three latent atmospheric levels. Separately, the surface fields are divided into P×P×V×TP \times P \times V \times T patches with V=4V = 4, embedded, and concatenated to the three latent atmospheric levels. The result is a latent state with dimensions H×W×4×512H \times W \times4 \times512, where HH is the number of latitudes divided by PP and WW is the number of longitudes divided by PP.

The latent state is the input to the backbone, which advances it by one 6 h6\,\mathrm{h} time step. The backbone’s U-Net architecture processes the latent state at several spatial resolutions and represents both small- and large-scale features. At the same time, the backbone propagates information across space and variables via stacked layers of 3D attention blocks. These are Shifted

a^{a}Prognostic fields are both inputs and outputs.

Window (Swin) Transformer blocks, which compute attention within local windows and shift the windows by a fixed offset between layers Liu et al. [18]. Local attention avoids the quadratic scaling of the cost of global attention with input size, and the shifted windows pass information between neighboring windows. The evolved latent space produced by the backbone is then passed to the decoder. The decoder learns to reverse the encoder’s operations to produce the predicted atmospheric and surface fields at t0+6ht_0 + 6h.

Training data

The chemical fields we use for fine-tuning come from a GEOS-Chem simulation for 2018–2019 performed for the FETCH4_4 project (Fate, Emissions, and Transport of CH4_4; https://fetch4.github.io/; last accessed 15 November 2026). GEOS-Chem is a global 3D chemical transport model driven by MERRA-2 meteorological fields that simulates the emissions, transport, chemistry, and deposition of more than 300 trace gases and aerosols [2]. This version (based on GEOS-Chem v14.5.3) adds radiocarbon monoxide (14^{14}CO) and methane isotopologues to the chemical mechanism. We output 3-hourly, time-averaged GEOS-Chem fields at a 4∘4^\circ by 5∘5^\circ horizontal resolution on 72 vertical levels following hybrid-sigma coordinates.

We fine-tune Aurora with seven chemical fields relevant for the atmosphere’s oxidative capacity (Table 1), and additional chemical fields can be easily added. OH and Cl are the primary chemical sinks of methane in the troposphere, accounting for approximately 88% and 1% of the methane sink [29]. NO and NO2_2 increase OH through the reaction NO + HO2_2. Ozone photolysis in the presence of water vapor is the primary source of OH [24]. CO is the largest single sink of OH [15]. 14^{14}CO is an emerging proxy for OH (e.g., [23]). Cosmic rays produce 14^{14}CO mainly at high latitudes and in the upper troposphere and lower stratosphere, and OH is its main sink in the troposphere. 14^{14}CO therefore may provide constraints on the OH seasonal cycle and interhemispheric gradient [26].

Category (Source)NamesLevelsRolea^{a}
Chemical (GEOS-Chem)OH, Cl, NO, NO2_2, O3_3, CO, 14^{14}CO3DPrognostic
Meteorological (ERA5)T, q, U, V, Z3DPrognostic
Surface meteorology (ERA5)2m T, 10m U, 10m V, mean sea level pressure2DPrognostic
Monthly mean emissions (CAMS)CO, NOx_x, NH3_3, SO2_22DInput
Static fieldsland-sea mask, soil type, topography, time of day, day of week, day of year2DInput

Table 1. Fields used to fine-tune AuroraChem.

The meteorological fields come from ERA5 [8], a state-of-the-art reanalysis produced by the European Center for Medium-Range Weather Forecasts (ECMWF). We use the same fields as base Aurora. Specifically, we use temperature, specific humidity, zonal and meridional wind, and geopotential height on pressure levels, and 2 m temperature, 10 m zonal and meridional wind, and mean sea level pressure at the surface (Table 1).

In addition to chemistry and climate fields, AuroraChem also ingests emissions, time, and static fields as inputs (Table 1).

The emissions are 2019 monthly mean anthropogenic emissions of CO, NOx_x, SO2_2, and NH3_3 from the CAMS global inventory (https://ads.atmosphere.copernicus.eu/datasets/cams-global-emission-inventories). We use the 2019 emissions for both the training (2018) and test (2019) years. AuroraChem receives no biomass burning or biogenic emissions. The time fields are time of day, day of week, and day of year. The static fields are land-sea mask, soil type, and topography.

We re-grid all fields to a 4∘4^\circ by 5∘5^\circ horizontal resolution and all 3D fields to the 13 isobaric pressure levels present in base Aurora: {1000,925,850,700,600,500,400,300,250,200,150,100,50}\{1000, 925, 850, 700, 600, 500, 400, 300, 250, 200, 150, 100, 50\} hPa. The 4∘×5∘4^\circ\times5^\circ resolution reflects the GEOS-Chem simulations available to us. For ERA5, we bilinearly interpolate the pressure-level and surface fields from 0.25∘×0.25∘0.25^\circ\times0.25^\circ to 4∘×5∘4^\circ\times5^\circ. For GEOS-Chem, we conservatively regrid each model level to the AuroraChem grid, then linearly interpolate each column from the hybrid sigma-pressure levels to the 13 pressure levels. For pressure levels greater than the surface pressure, we use the chemical concentration at the highest pressure level for the "below-ground" pressure levels. We bilinearly interpolate the 2D ERA5, CAMS emissions, and other static fields from their native resolutions to 4∘×5∘4^\circ\times5^\circ. We normalize each field at each pressure level before fine-tuning, following Bodnar et al. [4]. Meteorological fields are normalized by their mean and standard deviation (Eq. B5 in Bodnar et al., 2025). Chemical fields are normalized by half of their time-averaged maximum (Eq. B8 in Bodnar et al., 2025).

Fine-tuning AuroraChem

We fine-tune Aurora with GEOS-Chem chemical fields and ERA5 meteorological fields from January 1st through December 31st, 2018. Our fine-tuned model, AuroraChem (http://aurorachem.github.io), is then evaluated on chemical and meteorological fields that were unseen during training: January 1st through December 31st, 2019. We exclude 30 random samples from the training set to act as the validation set. We fine-tune Aurora for 3,000 steps with a batch size of 1 (60 epochs of 50 batches), about 2 passes through the 2018 data. Fine-tuning takes ∼2\sim2 h on one NVIDIA H100 GPU.

To accommodate additional chemical fields and a coarser spatial resolution, we make several changes to the base Aurora architecture. First, we reduce the patch size from 4×44 \times4 to 1×11 \times1. Base Aurora uses 4×44 \times4 patches to reduce cost at 0.25∘0.25^\circ resolution, but a patch of one grid cell is computationally affordable at 4∘×5∘4^\circ\times5^\circ. Second, we train a separate patch embedding for each pressure level, following the Aurora air quality model [4]. Chemical species vary more with altitude than meteorological fields do. Third, we train a new decoder for the chemical fields and keep the pretrained decoder for the meteorological fields. The separate decoder improves the accuracy of the chemical forecasts.

We update all weights during fine-tuning. The new encoder (7M parameters, 0.5% of the total) and decoder (18M parameters, 1.4%) are trained from scratch with a learning rate of 1×10−21 \times10^{-2}. The pretrained weights, including the backbone (1,240M parameters, 98%), are updated with a learning rate of 1×10−31 \times10^{-3}. The loss is a latitude-weighted mean absolute error with a separate weight for each field, following Bodnar et al. [4]. The weight for each chemical species is its normalization factor divided by the mean absolute error of a one-step (6 h) persistence forecast (Eq. D13 of Bodnar et al., 2025). This gives the forecast error of each field a similar contribution to the loss.

Free-running rollouts of chemistry and meteorology

Our ultimate goal is to conduct data assimilation on seasonal to annual timescales. The base Aurora air quality model was evaluated on 5-day forecasts [4], and ML emulators of chemistry have previously been shown to become unstable within weeks [11, 12]. This raises the question: “how long can AuroraChem maintain stable and realistic fields?”

We simulate year-long free-running AuroraChem rollouts. We initialize rollouts with the GEOS-Chem and ERA5 states at 00 and 06 UTC on 1 January 2019, and they receive no further information from the true fields. We note that AuroraChem was fine-tuned with data from 2018: all rollout evaluations are conducted on time stamps unseen during fine-tuning. We compare the rollouts with GEOS-Chem for the chemical fields and with ERA5 for the meteorological fields. A one-year rollout takes about 12 minutes on one NVIDIA H100 GPU. A one-year GEOS-Chem simulation takes about 4,000 minutes on a 32-core Intel Xeon Gold 6226R CPU albeit while simulating roughly 300 chemical species on 72 vertical levels compared to AuroraChem’s 7 chemical and 5 meteorological fields on 13 levels.

We evaluate the rollouts with the latitude-weighted spatial correlation coefficient between the predicted field PP and the true field TT at time tt:

Rs(t)=∑i,jw(i)(Pi,j(t)−P(t)‾)(Ti,j(t)−T(t)‾)∑i,jw(i)(Pi,j(t)−P(t)‾)2∑i,jw(i)(Ti,j(t)−T(t)‾)2,(2)R_s(t) = \frac{\sum_{i,j} w(i)\left(P_{i,j}(t)-\overline{P(t)}\right)\left(T_{i,j}(t)-\overline{T(t)}\right)}{\sqrt{\sum_{i,j} w(i)\left(P_{i,j}(t)-\overline{P(t)}\right)^2\sum_{i,j}w(i)\left(T_{i,j}(t)-\overline{T(t)}\right)^2}}, \tag*{(2)}
w(i)=cos⁡(lat⁡(i))1H∑k=1Hcos⁡(lat⁡(k))(3)w(i) = \frac{\cos(\operatorname{lat}(i))}{\frac{1}{H}\sum_{k=1}^{H}\cos(\operatorname{lat}(k))} \tag*{(3)}

where the sums are over the grid cells (i,j)(i,j) at one pressure level and the overbars denote means over those grid cells. RsR_s measures whether the spatial patterns of the predicted and true fields match. We weight each grid cell’s contribution to RsR_s by the cosine of its latitude to ensure that the metric is not biased towards performance at the poles. We also use latitude-weighted root mean squared error (RMSE) to evaluate rollouts (Supplemental Figures S3-6):

RMSE(t)=∑i=1H∑j=1Ww(i)(Pi,j(t)−Ti,j(t))2H∗W(4)RMSE(t) = \sqrt{\frac{\sum_{i=1}^{H}\sum_{j=1}^{W}w(i)\left(P_{i,j}(t)-T_{i,j}(t)\right)^2}{H * W}} \tag*{(4)}

At short lead times, AuroraChem retains the spatial patterns of both the meteorological and the chemical fields. Figure 2 compares AuroraChem forecasts at 3-day lead times with the true fields. The temperature forecast at 500 hPa has Rs=0.97R_s = 0.97 and retains the maxima over sub-Saharan Africa and Australia and the minima over Siberia and the Arctic (Figure 2b). The three chemical species displayed span a range of lifetimes: about 1 s for OH, about 3 weeks for ozone, and about 2 months for 14CO{}^{14}\mathrm{CO}. The OH forecast at 500 hPa has Rs=0.96R_s = 0.96 (Figure 2d). It has OH only in sunlit regions at 12 UTC and low OH over the Amazon, where OH is consumed by biogenic VOCs at the surface. The ozone forecast has Rs=0.91R_s = 0.91 and retains the interhemispheric gradient and the high ozone over India and sub-Saharan Africa from anthropogenic and biomass burning emissions (Figure 2f). The 14CO{}^{14}\mathrm{CO} forecast has Rs=0.96R_s = 0.96 and retains the interhemispheric gradient and the mid-latitude structure

AuroraChem 3-day forecast at 500 hPa

Figure 2. AuroraChem 3-day forecast at 500 hPa. Panels show output at 12 UTC on 4 January 2019. AuroraChem was initialized with the states at 00 and 06 UTC on 1 January 2019. The left column shows the true fields: ERA5 temperature (a) and GEOS-Chem OH (c), ozone (e), and 14CO{}^{14}\mathrm{CO} (g). The right column shows the AuroraChem forecasts (b, d, f, h). The value in each right-hand panel is RsR_s (Eq. (2)).

from advection (Figure 2h). However, it overestimates 14CO^{14}\mathrm{CO} globally. For short lead-times, AuroraChem forecasts maintain the characteristic global and regional spatial patterns for physical fields as well as chemical fields with varying lifetimes.

After six months, the AuroraChem rollout remains stable for the physical fields and short-lived species but develops spatial artifacts, particularly for fields without a strong seasonal cycle. Figure 3 shows the same fields as Figure 2 but for 4 July 2019 (6 months later). The temperature forecast keeps realistic values and large-scale patterns (Rs=0.90R_s = 0.90; Figure 3b). The OH forecast from AuroraChem similarly shows realistic concentrations and patterns (Rs=0.86R_s = 0.86; Figure 3d). It places the OH maximum at the correct longitude for 12 UTC, but it underestimates concentrations in the Northern Hemisphere, especially in the mid-latitudes. AuroraChem therefore gets the position of the OH maximum right for the time of day but does not fully capture the seasonal cycle. We do not observe the development of high-concentration artifacts in the spatial fields of short-lived species.

Global maps comparing True and AuroraChem forecasts for temperature, OH, ozone, and $^{14}\mathrm{CO}$

Figure 3. Same as Figure 2, but for the AuroraChem free rollout six months after initialization, valid at 12 UTC on 4 July 2019.

The longer-lived species develop artifacts, particularly in the tropics. The ozone forecast has Rs=0.59R_s = 0.59 (Figure 3f). It correctly estimates high ozone concentrations over the Middle East, Northern India, and China. However, it develops spatial artifacts over the Maritime Continent and over the Southern Atlantic. The 14CO{}^{14}\mathrm{CO} forecast has Rs=0.61R_s = 0.61 (Figure 3h). It has artifacts in the same tropical regions as ozone and misses the plume over Siberia in GEOS-Chem (Figure 3g). In GEOS-Chem, the 14CO{}^{14}\mathrm{CO} maximum moves from the Northern to the Southern Hemisphere during Northern Hemisphere summer, when OH is highest in the north. The six-month forecast correctly estimates elevated 14CO{}^{14}\mathrm{CO} in the Southern Hemisphere but also retains high concentrations in the Northern Hemisphere. Two factors contribute to these errors. First, AuroraChem receives only monthly mean anthropogenic emissions (§ 2.2). In a long-range simulation, it cannot represent transient emission events such as fires without external perturbations to the seasonal cycle, which produce features like the Siberian plume. Second, the 6 h change in a long-lived species is small compared with its spatial variability. The model must learn these small tendencies, and errors at each step accumulate over the rollout. AuroraChem rollouts of the meteorological fields and the short-lived species remain realistic for one year but the longer-lived species drift, and the rollouts do not fully capture the seasonal cycle. This raises the question of whether observations can restore the seasonal cycle and reduce the drift.

Idealized joint chemical and meteorological data assimilation

Joint chemical and meteorological data assimilation requires a forward model in which chemical observations update the meteorological state and meteorological observations update the chemical state. To test whether AuroraChem has learned relationships between chemistry and meteorology that allow this, we conduct idealized data assimilation experiments with direct insertion (DI) ([9]). In DI, the predicted fields are replaced with observed fields every NN time steps. We use the true fields as pseudo-observations and insert them every 5 days (N=20N = 20 time steps). Each insertion replaces both input states (t0−6t_0 - 6 h and t0t_0). As such, replacing both the chemical and meteorological fields is equivalent to re-initializing the model. We run three DI configurations for 2019: 1) full DI replaces all 3D fields, 2) met-only DI replaces only the 3D meteorological fields, and 3) chem-only DI replaces only the chemical fields. The 3D meteorological fields are T, q, U, V, and Z while the chemical fields are OH, Cl, NO, NO2_2, O3_3, CO, and 14CO{}^{14}\mathrm{CO}. The free rollout from § 3 serves as the control. We evaluate the rollouts with RsR_s (Eq. (2)) against GEOS-Chem for the chemical fields and ERA5 for the meteorological fields. All DI experiments are performed on 2019, which is out-of-sample.

Constraining global and zonal mean phenomena

Figure 4a shows RsR_s for 500 hPa temperature in each rollout. The free rollout keeps Rs≈0.9R_s \approx0.9 until May, then falls to ≈0.85\approx0.85 in June, which reflects the inability to fully capture the seasonal cycle, seen in Figure 3b. Full DI and met-only DI keep Rs≥0.9R_s \ge0.9 through the year, with annual means of 0.97 and 0.96, respectively. This is expected, because both insert the true temperature every 5 days. Chem-only DI never inserts temperature, yet it keeps Rs≈0.9R_s \approx0.9 through the Northern Hemisphere summer. Its annual mean RsR_s is 0.92, compared with 0.89 for the free rollout. Inserting the chemical fields alone is therefore enough to keep the temperature field in the correct season.

Spatial pattern correlations for AuroraChem 2019 rollouts

Figure 4. Time series of spatial pattern correlations of AuroraChem rollouts compared to the true fields. Panels show RsR_s (Eq. (2)) at 500 hPa for (a) temperature, (b) OH, and (c) ozone. Lines show the free rollout (dashed), full DI (black), met-only DI (blue), and chem-only DI (orange). Values in the legends are annual means. Fields are inserted every 5 days, which produces the sawtooth pattern.

Figure 4b shows the same comparison for 500 hPa OH. The free rollout keeps Rs≈0.90R_s \approx0.90 until mid-April, then falls to ≈0.80\approx0.80 in mid-June. Full DI and chem-only DI keep Rs>0.90R_s > 0.90 for most of the year, with annual means of 0.96 and 0.94. This is expected, because both insert the true OH. Similar to the ability of chem-only DI to recover temperature, we find that met-only DI is also able to recover OH with an annual mean RsR_s of 0.93, close to that of full DI and chem-only DI. OH depends on photolysis and water vapor (Eq. (1)), and the meteorological fields carry information about both. As such, inserting only the meteorological fields constrains the OH spatial pattern nearly as well as inserting OH and its chemical drivers. The radiative and dynamical information encoded in meteorological fields is sufficient to constrain the OH spatial pattern.

Ozone rollouts with met-only direct insertion further demonstrate that AuroraChem has learned realistic climate-to-chemistry relationships. Figure 4c shows the results for 500 hPa ozone. The AuroraChem 500 hPa ozone free rollout develops spatial artifacts at long lead-times (Figure 3f); this can be seen in the free rollout decline through the year to Rs≈0.40R_s \approx0.40 in December. Full DI and chem-only DI keep Rs≈0.80R_s \approx0.80 until October, with annual means of 0.88 and 0.82. In both, RsR_s drops from near 1 at each insertion to about 0.7 five days later, indicating that ozone errors grow quickly even from the true state. Met-only DI maintains RsR_s similar to that of full and chem-only DI five days after each insertion, with an annual mean of 0.72. This demonstrates that AuroraChem has learned sufficient climate-to-chemistry relationships to improve the spatial realism of ozone forecasts using information encoded in meteorological fields, resulting in an ozone field about as accurate as a 5-day forecast from the true ozone.

In Figure 5, we further demonstrate the ability of each direct insertion configuration to recover the OH seasonal cycle. Figure 5 shows the weekly zonal mean 500 hPa OH through 2019. In GEOS-Chem, OH peaks in the subtropics of the summer hemisphere and is lowest over the winter pole (Figure 5a). The Northern Hemisphere summer maximum exceeds the Southern Hemisphere maximum because of higher ozone and NOx\mathrm{NO}_x in the north in GEOS-Chem. The free rollout has a weak OH seasonal cycle and does not represent the magnitude of the subtropical summer maximum (Figure 5b). All three DI rollouts better capture the timing and magnitude of the seasonal cycle (Figure 5c–Figure 5e). Met-only DI, which never inserts OH, captures the subtropical summer maxima, the polar winter minima, and the larger Northern Hemisphere maximum, though it underestimates polar summertime OH compared to GEOS-Chem (Figure 5d). Inserting only the meteorological fields is able to recover the main spatiotemporal features of the OH seasonal cycle.

Hovmöller plots of weekly averaged zonal mean OH at 500 hPa

Figure 5. Hovmöller plots of weekly averaged zonal mean OH at 500 hPa from (a) GEOS-Chem and from AuroraChem rollouts with (b) no DI, (c) full DI, (d) met-only DI, and (e) chem-only DI. The dashed line marks the equator.

Most of our evaluations thus far have focused on the free troposphere (500 hPa). Figure 6 shows the 2019 zonal mean OH at all 13 pressure levels. The free rollout captures the general vertical and zonal structure, with higher concentrations in the mid-troposphere over the tropics and lower concentrations near the surface over the equator. However, the free rollout underestimates OH in the northern subtropics and in the upper troposphere over the mid-latitudes (Figure 6b). All three DI rollouts better capture the vertical and latitudinal structure of the GEOS-Chem OH (Figure 6c–Figure 6e). The met-only DI rollout estimates greater lower troposphere OH in the Northern Hemisphere than the Southern Hemisphere, better aligning with the GEOS-Chem [image ends]

Annual zonal and vertical mean OH from GEOS-Chem and AuroraChem rollouts

Figure 6. Annual zonal mean OH as a function of latitude and pressure from (a) GEOS-Chem and from AuroraChem rollouts with (b) no DI, (c) full DI, (d) met-only DI, and (e) chem-only DI. The dashed line marks the equator.

Chem inter-hemispheric OH gradient (Figure 6d). The relationships between chemistry and meteorology that AuroraChem has learned therefore hold throughout the troposphere, not only at 500 hPa.

Recovering physical and dynamical features of regional phenomena

We have demonstrated that meteorological and chemical data can constrain each other at global and zonal mean scales. However, the Indian summer monsoon, a regional circulation that depends on the season, is a more demanding test than global or zonal means. The Indian summer monsoon brings more than 85% of India’s annual rainfall [36], and the free rollout, which stays in Northern Hemisphere winter, does not capture the dynamical characteristics of the monsoon. This raises the question of whether chemical observations alone can recover the Indian summer monsoon. In ERA5, the July mean 850 hPa specific humidity is high over India and Southeast Asia, and a low-level southwesterly jet, the Somali jet, crosses the Arabian Sea from the coast of Somalia to India (Figure 7a, Figure 7d). The free rollout places the moisture loading to the west of the ERA5 field, and the southwesterly jet is much weaker and to the north of the ERA5 winds (Figure 7b, Figure 7e). Chem-only DI never inserts humidity or wind, yet it recovers both the high humidity over India and the magnitude and position of the southwesterly jet (Figure 7c, Figure 7f). However, the chem-only DI rollout overestimates specific humidity over the Arabian Sea, and its jet is slightly weaker and wider than in ERA5. Inserting the chemical fields every 5 days is therefore sufficient to recover the main moisture and circulation features of the monsoon.

July 2019 Indian Summer Monsoon: specific humidity and wind speed at 850 hPa shown for ERA5 (Truth), Free Rollout, and Chem DI

Figure 7. July 2019 mean (a–c) specific humidity and (d–f) wind speed at 850 hPa from (a, d) ERA5 and from AuroraChem rollouts with (b, e) no DI and (c, f) chem-only DI. Arrows show the wind direction and magnitude.

Constraining AuroraChem with TROPOMI-like NO2_2 observations

Thus far, we have assumed full spatial coverage in our hypothetical observing system. However, real observing systems are sparse in space, time, or both. Here we investigate the impact of a more realistic observing strategy on the ability of AuroraChem to recover chemical or meteorological information using an observing system representative of NO2_2 retrievals from the TROPOspheric Monitoring Instrument (TROPOMI) on Sentinel-5P [34].

TROPOMI retrieves NO2NO_2 columns daily (13:30 local time) with near-global coverage, but clouds and retrieval failures leave gaps, and each retrieval measures a column whose vertical sensitivity is described by an averaging kernel. We run three DI experiments, each successively closer to assimilating a real TROPOMI NO2NO_2 retrieval to assess the loss of information from each step: 1) NO2NO_2-only DI inserts the true 3D NO2NO_2 field everywhere, as in § 4, but inserts every time-step and only NO2NO_2 rather than all seven chemical species, 2) sampled NO2NO_2 DI inserts the true NO2NO_2 profile every time-step, but only in grid cells that contain valid TROPOMI pixels within ±3\pm3 hours, and 3) sampled NO2NO_2 column DI uses the same grid cells as 2) but inserts only the column. The final test uniformly scales the predicted NO2NO_2 profile in each cell by the ratio of the true to the predicted NO2NO_2 column, with both columns computed using a globally representative TROPOMI NO2NO_2 averaging kernel. Therefore, true vertical profile NO2NO_2 information is not directly inserted. For the sampled experiments, we use all 2019 reprocessed TROPOMI Level 2 NO2NO_2 retrievals with quality assurance values above 0.75, following Sekiya et al. [30], and we interpolate the averaging kernels from the retrieval coordinates to the AuroraChem grid.

Spatial pattern correlations for AuroraChem 2019 rollouts

Figure 8. RsR_s (Eq. (2)) at 500 hPa for the 2019 AuroraChem rollouts with TROPOMI-like NO2_2 DI for (a) temperature, (b) OH, and (c) ozone. Lines show the free rollout (dashed), full DI (black), NO2_2-only DI (dark green), sampled NO2_2 DI (medium green), and sampled NO2_2 column DI (light green). Values in the legends are annual means.

fig:8a shows RsR_s for 500 hPa temperature in the NO2_2 DI experiments. All three NO2_2 DI experiments have an annual mean temperature Rs=0.90R_s = 0.90 to 0.910.91, comparable to the free rollout RsR_s of 0.890.89. NO2_2 information alone is not sufficient to recover the seasonal cycle of temperature. fig:8b shows the 500 hPa OH RsR_s of NO2_2 DI experiments. NO2_2-only DI keeps OH Rs≥0.90R_s \ge0.90 through 2019, with an annual mean of 0.950.95, and the sampled and column experiments both have annual means of 0.930.93, compared with 0.890.89 for the free rollout. Daily, sparse NO2_2 column information constrains OH nearly as well as complete 3D NO2_2 fields.

fig:8c shows the NO2_2 DI results for 500 hPa ozone. NO2_2-only DI raises the annual mean RsR_s from 0.550.55 in the free rollout to 0.690.69, comparable to the RsR_s value full DI reaches five days after each insertion. Sampling at TROPOMI pixels lowers the annual mean RsR_s to 0.620.62, even with daily insertion, and inserting only column information lowers RsR_s to 0.590.59, still above the free rollout. Ozone is more sensitive than temperature or OH to how NO2_2 is inserted. Sparse NO2_2 column observations therefore constrain OH, partly constrain ozone, and do not constrain temperature in AuroraChem rollouts. It is important to caveat that these experiments use noise-free pseudo-observations, so they test the effects of sparse sampling and column integration but not the effect of retrieval errors.

What is AuroraChem learning?

We have shown that inserting the chemical fields constrains the meteorological state in AuroraChem and inserting the meteorological fields constrains the chemical state. This raises the question “Is AuroraChem learning the physics that links chemistry and meteorology, or just the correlations between them?” We posit here that AuroraChem is primarily learning the correlation structure, but that this is still sufficient for the purposes of data assimilation. Some of the relationships we find are consistent with the underlying physics. For example, met-only DI recovers the OH seasonal cycle, as we would expect from the dependence of OH on photolysis and water vapor (Eq. (1)). However, other relationships cannot be physical. Chem-only DI recovers the moisture and the Somali jet of the Indian summer monsoon, yet chemistry does not drive the monsoon on 5-day time scales, and the training data contain no feedback from chemistry to meteorology because GEOS-Chem is driven by prescribed meteorology. Instead, the chemical fields carry an imprint of the meteorology through transport, photolysis, and the seasonal cycle, and AuroraChem uses this imprint to infer the meteorological state. As such, it is likely that AuroraChem has learned the correlation structure between the chemical and meteorological fields, not necessarily the physics that produces it.

Consider the analogous problem of keeping a car on the road through a turn. There is a lot of physics behind this that depends on the conditions (e.g., speed, tire friction, ice on the road), but drivers do not solve these equations. Instead, they learn from experience which steering input corresponds to which view of the road. These learned correlations work well on roads similar to those the driver has seen before but fail on ice. We expect AuroraChem to behave similarly when perturbed outside of its training data. For example, AuroraChem was trained with the monthly mean emissions for one year, so it has never seen how chemistry responds to inter-annual changes in emissions, and it is unlikely to produce the correct OH response to a perturbation in NOx_x emissions. This means AuroraChem is not well suited for perturbation experiments that attribute changes in OH to their underlying drivers. Nevertheless, AuroraChem is well suited to some data assimilation applications. Data assimilation requires knowing how an observation of one variable should update the others, and this is determined by the correlation structure. The goal of joint chemical and meteorological data assimilation is to keep OH consistent with the observations, much like the goal of the driver is to keep the car on the road, and the correlations learned by AuroraChem are sufficient for this purpose.

Conclusions

Constraining OH requires a model that can assimilate both chemical and meteorological observations. Coupled chemistry-climate models are presently too computationally expensive to use for this task. In this work, we present AuroraChem (http://aurorachem.github.io), a fine-tuned version of the Aurora atmospheric foundation model with the chemical fields needed to constrain OH. Fine-tuning Aurora on one year of chemical and meteorological fields produces a model (AuroraChem) that simulates a one-year rollout in about 12 minutes on one GPU. Free rollouts from AuroraChem remain stable for one year but do not fully capture seasonal cycles, and the longer-lived species develop artifacts in the tropics. Using an idealized data assimilation framework, we find that assimilating chemical fields allows the model to recover the seasonal cycle of the meteorology, and assimilating meteorological fields recovers the seasonal cycle of OH. Assimilating only NO2NO_2, sampled daily like TROPOMI, constrains OH nearly as well as inserting complete NO2NO_2 fields. These results show that a fine-tuned atmospheric foundation model provides a new pathway to joint chemical and meteorological data assimilation.

AuroraChem builds on a pretrained foundation model rather than an emulator trained from scratch. We fine-tune Aurora with seven chemical species from a GEOS-Chem simulation (OH, Cl, NO, NO2NO_2, ozone, CO, and 14CO{}^{14}\mathrm{CO}) and five meteorological fields from ERA5, regridded to 4∘×5∘4^\circ\times5^\circ and 13 pressure levels. Fine-tuning requires three changes to the architecture: a patch size of one grid cell, a separate patch embedding for each pressure level, and a separate decoder for the chemical fields. The pretrained backbone holds 98% of the 1.3 billion parameters and is updated at a low learning rate, so AuroraChem retains most of the pretrained model. A one-year AuroraChem rollout takes about 12 minutes on one H100 GPU, compared with about 4,000 minutes for a one-year GEOS-Chem simulation on a 32-core CPU.

We evaluate AuroraChem with free rollouts initialized on 1 January 2019. After three days, the 500 hPa forecasts retain the spatial patterns of temperature, OH, ozone, and 14CO{}^{14}\mathrm{CO}, with RsR_s between 0.91 and 0.97. After six months, temperature and OH keep realistic values and spatial patterns (RsR_s of 0.90 and 0.86, respectively), but the rollout state retains some features of Northern Hemisphere winter. OH peaks at the correct longitude for the time of day but does not fully capture the northward shift in the maxima during Northern Hemisphere summertime. The longer-lived species drift further, with RsR_s of 0.59 for ozone and 0.61 for 14CO{}^{14}\mathrm{CO}, and develop artifacts over the Amazon, the Congo basin, and the Maritime Continent. These errors likely reflect the single year of training data, the absence of transient emissions such as fires, and the small 6 h tendencies of long-lived species.

We use direct insertion to test whether observations of chemistry constrain meteorology, and vice versa. Inserting only the chemical fields every 5 days keeps 500 hPa temperature at an annual mean RsR_s of 0.92, compared with 0.89 for the free rollout, and recovers the moisture loading and the southwesterly Somali jet of the Indian summer monsoon. Inserting only the meteorological fields keeps 500 hPa OH at 0.93, compared with 0.89 for the free rollout and 0.94 for inserting the chemical fields themselves. Inserting meteorology also recovers the latitudinal and vertical structure of the OH seasonal cycle and raises ozone RsR_s from 0.55 to 0.72. These relationships are statistical. AuroraChem has likely learned covariances between the chemical and meteorological fields rather than the physical coupling between them. However, these covariances are what data assimilation uses to let an observation of one variable update the others.

We then test whether sparse, column-integrated observations of a single species can constrain AuroraChem. Inserting only NO2\mathrm{NO}_2 at every time-step keeps OH at an annual mean RsR_s of 0.95. Inserting NO2\mathrm{NO}_2 1) only in grid cells with valid TROPOMI pixels and 2) only as columns computed with a globally representative TROPOMI averaging kernel, gives annual mean OH RsR_s of 0.93 for both. Ozone is more sensitive to the NO2\mathrm{NO}_2 sampling. Its RsR_s falls from 0.69 with complete NO2\mathrm{NO}_2 fields to 0.59 with TROPOMI-like columns, though it stays above the free rollout (0.55). Sparse NO2\mathrm{NO}_2 column observations can therefore constrain OH in AuroraChem.

AuroraChem could serve as the forward model for joint assimilation of chemical and meteorological observations to constrain OH. Its low cost makes ensemble data assimilation practical. Its learned covariances can inform the cross-variable terms of the background error covariance, which are difficult to specify by hand and have limited the application of data assimilation to chemistry-climate models (e.g., Bocquet et al., [3]). Because it is differentiable, it provides the gradients for 4D variational data assimilation without a hand-coded adjoint model (e.g., Vonich and Hakim, [35]). Several areas for improvement remain. Fine-tuning on multiple years of chemistry-climate model output should improve the seasonal cycle and add the feedbacks from chemistry to meteorology that GEOS-Chem lacks. Fine-tuning at 1∘×1∘1^\circ\times1^\circ would better resolve short-lived species such as OH, and adding CH4\mathrm{CH}_4 and its isotopologues (13CH4{}^{13}\mathrm{CH}_4, 12CH3D{}^{12}\mathrm{CH}_3\mathrm{D}, and 13CH3D{}^{13}\mathrm{CH}_3\mathrm{D}) would allow assimilation of methane observations. The next step is to assimilate real meteorological and chemical observations, including TROPOMI NO2\mathrm{NO}_2, to estimate three-dimensional, time-resolved OH. Such estimates would constrain recent changes in OH, and its role in the methane budget, independently of methyl chloroform.

Code availability. The code for this study is available at https://github.com/aurorachem/AuroraChem (last access: 16 November 2026) and code will be uploaded to Zenodo upon acceptance.

Author contributions. E.J.K. and A.J.T. designed research; E.J.K. and A.J.T. performed research; E.J.K., E.J.M., P.D.I., L.T.M., and A.J.T. analyzed data; and E.J.K., E.J.M., P.D.I., L.T.M., and A.J.T. wrote the paper.

Competing interests. The authors do not declare any competing interests.

Acknowledgements. This work is part of the FETCH4 project, and support was provided by Schmidt Sciences. E.J.M. acknowledges funding from a National Defense Science and Engineering Graduate Fellowship from the Department of Defense, supported by the Air Force Office of Scientific Research under award number FA9550-25-C-B010 in the dollar amount of $143,000.

References

  1. [1]Anderson, D. C., Duncan, B. N., Liu, J., Nicely, J. M., Strode, S. A., Follette-Cook, M. B., Souri, A. H., Ziemke, J. R., González-Abad, G., and Ayazpour, Z.: Trends and Interannual Variability of the Hydroxyl Radical in the Remote Tropics During Boreal Autumn Inferred From Satellite Proxy Data, Geophysical Research Letters, 51, e2024GL108531, https://doi.org/10.1029/2024GL108531, _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1029/2024GL108531, 2024.
  2. [2]Bey, I., Jacob, D. J., Yantosca, R. M., Logan, J. A., Field, B. D., Fiore, A. M., Li, Q., Liu, H. Y., Mickley, L. J., and Schultz, M. G.: Global modeling of tropospheric chemistry with assimilated meteorology: Model description and evaluation, Journal of Geophysical Research: Atmospheres, 106, 23 073–23 095, https://doi.org/10.1029/2001JD000807, _eprint: https://agupubs.onlinelibrary.wiley.com/doi/pdf/10.1029/2001JD000807, 2001.
  3. [3]Bocquet, M., Elbern, H., Eskes, H., Hirtl, M., Žabkar, R., Carmichael, G. R., Flemming, J., Inness, A., Pagowski, M., Pérez Camaño, J. L., Saide, P. E., San Jose, R., Sofiev, M., Vira, J., Baklanov, A., Carnevale, C., Grell, G., and Seigneur, C.: Data assimilation in atmospheric chemistry models: current status and future prospects for coupled chemistry meteorology models, Atmospheric Chemistry and Physics, 15, 5325–5358, https://doi.org/10.5194/acp-15-5325-2015, 2015.
  4. [4]Bodnar, C., Bruinsma, W. P., Lucic, A., Stanley, M., Allen, A., Brandstetter, J., Garvan, P., Riechert, M., Weyn, J. A., Dong, H., Gupta, J. K., Thambiratnam, K., Archibald, A. T., Wu, C.-C., Heider, E., Welling, M., Turner, R. E., and Perdikaris, P.: A foundation model for the Earth system, Nature, 641, 1180–1187, https://doi.org/10.1038/s41586-025-09005-y, 2025.
  5. [5]Bonev, B., Kurth, T., Mahesh, A., Bisson, M., Kossaifi, J., Kashinath, K., Anandkumar, A., Collins, W. D., Pritchard, M. S., and Keller, A.: FourCastNet 3: A geometric approach to probabilistic machine-learning weather forecasting at scale, https://doi.org/10.48550/arXiv.2507.12144, arXiv:2507.12144 [cs], 2025.
  6. [6]Cresswell-Clay, N., Liu, B., Durran, D. R., Liu, Z., Espinosa, Z. I., Moreno, R. A., and Karlbauer, M.: A Deep Learning Earth System Model for Efficient Simulation of the Observed Climate, AGU Advances, 6, e2025AV001706, https://doi.org/10.1029/2025AV001706, 2025.
  7. [7]Emmons, L. K., Schwantes, R. H., Orlando, J. J., Tyndall, G., Kinnison, D., Lamarque, J.-F., Marsh, D., Mills, M. J., Tilmes, S., Bardeen, C., Buchholz, R. R., Conley, A., Gettelman, A., Garcia, R., Simpson, I., Blake, D. R., Meinardi, S., and Pétron, G.: The Chemistry Mechanism in the Community Earth System Model Version 2 (CESM2), Journal of Advances in Modeling Earth Systems, 12, e2019MS001882, https://doi.org/10.1029/2019MS001882, _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1029/2019MS001882, 2020.
  8. [8]Hersbach, H., Bell, B., Berrisford, P., Hirahara, S., Horányi, A., Muñoz-Sabater, J., Nicolas, J., Peubey, C., Radu, R., Schepers, D., Simmons, A., Soci, C., Abdalla, S., Abellan, X., Balsamo, G., Bechtold, P., Biavati, G., Bidlot, J., Bonavita, M., De Chiara, G., Dahlgren, P., Dee, D., Diamantakis, M., Dragani, R., Flemming, J., Forbes, R., Fuentes, M., Geer, A., Haimberger, L., Healy, S., Hogan, R. J., Hólm, E., Janisková, M., Keeley, S., Laloyaux, P., Lopez, P., Lupu, C., Radnoti, G., de Rosnay, P., Rozum, I., Vamborg, F., Villaume, S., and Thépaut, J.-N.: The ERA5 global reanalysis, Quarterly Journal of the Royal Meteorological Society, 146, 1999–2049, https://doi.org/10.1002/qj.3803, 2020.
  9. [9]Houser, P. R., Shuttleworth, W. J., Famiglietti, J. S., Gupta, H. V., Syed, K. H., and Goodrich, D. C.: Integration of soil moisture remote sensing and hydrologic modeling using data assimilation, Water Resources Research, 34, 3405–3420, https://doi.org/10.1029/1998WR900001, 1998.
  10. [10]Houtekamer, P. L. and Zhang, F.: Review of the Ensemble Kalman Filter for Atmospheric Data Assimilation, Monthly Weather Review, 144, 4489–4532, https://doi.org/10.1175/MWR-D-15-0440.1, 2016.
  11. [11]Keller, C. A. and Evans, M. J.: Application of random forest regression to the calculation of gas-phase chemistry within the GEOS-Chem chemistry model v10, Geoscientific Model Development, 12, 1209–1225, https://doi.org/10.5194/gmd-12-1209-2019, 2019.
  12. [12]Kelp, M. M., Jacob, D. J., Kutz, J. N., Marshall, J. D., and Tessum, C. W.: Toward Stable, General Machine-Learned Models of the Atmospheric Chemical System, Journal of Geophysical Research: Atmospheres, 125, e2020JD032759, https://doi.org/10.1029/2020JD032759, _eprint: https://agupubs.onlinelibrary.wiley.com/doi/pdf/10.1029/2020JD032759, 2020.
  13. [13]Kochkov, D., Yuval, J., Langmore, I., Norgaard, P., Smith, J., Mooers, G., Klöwer, M., Lottes, J., Rasp, S., Düben, P., Hatfield, S., Battaglia, P., Sanchez-Gonzalez, A., Willson, M., Brenner, M. P., and Hoyer, S.: Neural general circulation models for weather and climate, Nature, 632, 1060–1066, https://doi.org/10.1038/s41586-024-07744-y, 2024.
  14. [14]Lam, R., Sanchez-Gonzalez, A., Willson, M., Wirnsberger, P., Fortunato, M., Alef, F., Ravuri, S., Ewalds, T., Eaton-Rosen, Z., Hu, W., Merose, A., Hoyer, S., Holland, G., Vinyals, O., Stott, J., Pritzel, A., Mohamed, S., and Battaglia, P.: Learning skillful medium-range global weather forecasting, Science, 382, 1416–1421, https://doi.org/10.1126/science.adi2336, 2023.
  15. [15]Lelieveld, J., Gromov, S., Pozzer, A., and Taraborrelli, D.: Global tropospheric hydroxyl distribution, budget and reactivity, Atmospheric Chemistry and Physics, 16, 12 477–12 493, https://doi.org/10.5194/acp-16-12477-2016, 2016.
  16. [16]Liang, Q., Chipperfield, M. P., Fleming, E. L., Abraham, N. L., Braesicke, P., Burkholder, J. B., Daniel, J. S., Dhomse, S., Fraser, P. J., Hardiman, S. C., Jackman, C. H., Kinnison, D. E., Krummel, P. B., Montzka, S. A., Morgenstern, O., McCulloch, A., Mühle, J., Newman, P. A., Orkin, V. L., Pitari, G., Prinn, R. G., Rigby, M., Rozanov, E., Stenke, A., Tummon, F., Velders, G. J. M., Visioni, D., and Weiss, R. F.: Deriving Global OH Abundance and Atmospheric Lifetimes for Long-Lived Gases: A Search for CH3CCl3 Alternatives, Journal of Geophysical Research: Atmospheres, 122, 11,914–11,933, https://doi.org/10.1002/2017JD026926, _eprint: https://agupubs.onlinelibrary.wiley.com/doi/pdf/10.1002/2017JD026926, 2017.
  17. [17]Liu, X., Mizzi, A. P., Anderson, J. L., Fung, I., and Cohen, R. C.: The potential for geostationary remote sensing of NO₂ to improve weather prediction, Atmospheric Chemistry and Physics, 21, 9573–9583, https://doi.org/10.5194/acp-21-9573-2021, 2021a.
  18. [18]Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., and Guo, B.: Swin Transformer: Hierarchical Vision Transformer using Shifted Windows, in: 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 9992–10 002, ISSN 2380-7504, https://doi.org/10.1109/ICCV48922.2021.00986, iSSN: 2380-7504, 2021b.
  19. [19]Mei, E. J., Hakim, G. J., Taniguchi-King, M., Stiller, D., and Turner, A. J.: Emulating chemistry-climate dynamics with a linear inverse model, Atmospheric Chemistry and Physics, 25, 15 033–15 045, https://doi.org/10.5194/acp-25-15033-2025, 2025.
  20. [20]Miyazaki, K., Bowman, K. W., Yumimoto, K., Walker, T., and Sudo, K.: Evaluation of a multi-model, multi-constituent assimilation framework for tropospheric chemical reanalysis, Atmospheric Chemistry and Physics, 20, 931–967, https://doi.org/10.5194/acp-20-931-2020, 2020.
  21. [21]Moldovan, G., Pinnington, E., Prieto Nemesio, A., Lang, S., Ben Bouallègue, Z., Dramsch, J., Alexe, M., Santa Cruz, M., Hahner, S., Cook, H., Theissen, H., Clare, M., O’Brien, C., Polster, J., Magnusson, L., Mertes, G., Pinault, F., Raoult, B., de Rosnay, P., Forbes, R., and Chantry, M.: AIFS Single 1.1.0: an update to ECMWF’s machine-learned weather forecast model AIFS, Geoscientific Model Development, 19, 4703–4724, https://doi.org/10.5194/gmd-19-4703-2026, 2026.
  22. [23]Morgenstern, O., Moss, R., Manning, M., Zeng, G., Schaefer, H., Usoskin, I., Turnbull, J., Brailsford, G., Nichol, S., and Bromley, T.: Radio-carbon monoxide indicates increasing atmospheric oxidizing capacity, Nature Communications, 16, 249, https://doi.org/10.1038/s41467-024-55603-1, 2025.
  23. [24]Murray, L. T., Mickley, L. J., Kaplan, J. O., Sofen, E. D., Pfeiffer, M., and Alexander, B.: Factors controlling variability in the oxidative capacity of the troposphere since the Last Glacial Maximum, Atmospheric Chemistry and Physics, 14, 3589–3622, https://doi.org/10.5194/acp-14-3589-2014, 2014.
  24. [25]Murray, L. T., Fiore, A. M., Shindell, D. T., Naik, V., and Horowitz, L. W.: Large uncertainties in global hydroxyl projections tied to fate of reactive nitrogen and carbon, Proceedings of the National Academy of Sciences, 118, e2115204118, https://doi.org/10.1073/pnas.2115204118, 2021.
  25. [26]Petrenko, V. V., Smith, A. M., Crosier, E. M., Kazemi, R., Place, P., Colton, A., Yang, B., Hua, Q., and Murray, L. T.: An improved method for atmospheric ¹⁴CO measurements, Atmospheric Measurement Techniques, 14, 2055–2063, https://doi.org/10.5194/amt-14-2055-2021, 2021.
  26. [27]Rigby, M., Montzka, S. A., Prinn, R. G., White, J. W. C., Young, D., O’Doherty, S., Lunt, M. F., Ganesan, A. L., Manning, A. J., Simmonds, P. G., Salameh, P. K., Harth, C. M., Mühle, J., Weiss, R. F., Fraser, P. J., Steele, L. P., Krummel, P. B., McCulloch, A., and Park, S.: Role of atmospheric oxidation in recent methane growth, Proceedings of the National Academy of Sciences, 114, 5373–5377, https://doi.org/10.1073/pnas.1616426114, 2017.
  27. [28]Saide, P. E., Spak, S. N., Pierce, R. B., Otkin, J. A., Schaack, T. K., Heidinger, A. K., Silva, A. M. d., Kacenelenbogen, M., Redemann, J., and Carmichael, G. R.: Central American biomass burning smoke can increase tornado severity in the U.S., Geophysical Research Letters, 42, 956–965, https://doi.org/10.1002/2014GL062826, 2015.
  28. [29]Saunois, M., Martinez, A., Poulter, B., Zhang, Z., Raymond, P. A., Regnier, P., Canadell, J. G., Jackson, R. B., Patra, P. K., Bousquet, P., Ciais, P., Dlugokencky, E. J., Lan, X., Allen, G. H., Bastviken, D., Beerl_ing, D. J., Belikov, D. A., Blake, D. R., Castaldi, S., Crippa, M., Deemer, B. R., Dennison, F., Etiope, G., Gedney, N., Höglund-Isaksson, L., Holgerson, M. A., Hopcroft, P. O., Hugelius, G., Ito, A., Jain, A. K., Janardanan, R., Johnson, M. S., Kleinen, T., Krummel, P. B., Lauerwald, R., Li, T., Liu, X., McDonald, K. C., Melton, J. R., Mühle, J., Müller, J., Murguia-Flores, F., Niwa, Y., Noce, S., Pan, S., Parker, R. J., Peng, C., Ramonet, M., Riley, W. J., Rocher-Ros, G., Rosentreter, J. A., Sasakawa, M., Segers, A., Smith, S. J., Stanley, E. H., Thanwerdas, J., Tian, H., Tsuruta, A., Tubiello, F. N., Weber, T. S., van der Werf, G. R., Worthy, D. E. J., Xi, Y., Yoshida, Y., Zhang, W., Zheng, B., Zhu, Q., Zhu, Q., and Zhuang, Q.: Global Methane Budget 2000–2020, Earth System Science Data, 17, 1873–1958, https://doi.org/10.5194/essd-17-1873-2025, 2025.
  29. [30]Sekiya, T., Miyazaki, K., Eskes, H., Sudo, K., Takigawa, M., and Kanaya, Y.: A comparison of the impact of TROPOMI and OMI tropospheric NO₂ on global chemical data assimilation, Atmospheric Measurement Techniques, 15, 1703–1728, https://doi.org/10.5194/amt-15-1703-2022, 2022.
  30. [31]Stone, D., Whalley, L. K., and Heard, D. E.: Tropospheric OH and HO2 radicals: field measurements and model comparisons, Chemical Society Reviews, 41, 6348–6404, https://doi.org/10.1039/C2CS35140D, 2012.
  31. [32]Turner, A. J., Frankenberg, C., Wennberg, P. O., and Jacob, D. J.: Ambiguity in the causes for decadal trends in atmospheric methane and hydroxyl, Proceedings of the National Academy of Sciences, 114, 5367–5372, https://doi.org/10.1073/pnas.1616020114, 2017.
  32. [33]Turner, A. J., Fung, I., Naik, V., Horowitz, L. W., and Cohen, R. C.: Modulation of hydroxyl variability by ENSO in the absence of external forcing, Proceedings of the National Academy of Sciences, 115, 8931–8936, https://doi.org/10.1073/pnas.1807532115, 2018.
  33. [34]Veefkind, J. P., Aben, I., McMullan, K., Förster, H., de Vries, J., Otter, G., Claas, J., Eskes, H. J., de Haan, J. F., Kleipool, Q., van Weele, M., Hasekamp, O., Hoogeven, R., Landgraf, J., Snel, R., Tol, P., Ingmann, P., Voors, R., Kruizinga, B., Vink, R., Visser, H., and Levelt, P. F.: TROPOMI on the ESA Sentinel-5 Precursor: A GMES mission for global observations of the atmospheric composition for climate, air quality and ozone layer applications, Remote Sensing of Environment, 120, 70–83, https://doi.org/10.1016/j.rse.2011.09.027, 2012.
  34. [35]Vonich, P. T. and Hakim, G. J.: Predictability Limit of the 2021 Pacific Northwest Heatwave From Deep-Learning Sensitivity Analysis, Geophysical Research Letters, 2024.arxiv.org/abs/2406.05019
  35. [36]Wang, B. and LinHo: Rainy Season of the Asian–Pacific Summer Monsoon, Journal of Climate, 15, 386–398, https://doi.org/10.1175/1520-0442(2002)015<0386:RSOTAP>2.0.CO;2, 2002.
  36. [37]Watt-Meyer, O., Henn, B., McGibbon, J., Clark, S. K., Kwa, A., Perkins, W. A., Wu, E., Harris, L., and Bretherton, C. S.: ACE2: accurately learning subseasonal to decadal atmospheric variability and forced responses, npj Climate and Atmospheric Science, 8, 205, https://doi.org/10.1038/s41612-025-01090-0, 2025.
  37. [38]Wolfe, G. M., Nicely, J. M., St. Clair, J. M., Hanisco, T. F., Liao, J., Oman, L. D., Brune, W. B., Miller, D., Thames, A., González Abad, G., Ryerson, T. B., Thompson, C. R., Peischl, J., McKain, K., Sweeney, C., Wennberg, P. O., Kim, M., Crounse, J. D., Hall, S. R., Ullmann, K., Diskin, G., Bui, P., Chang, C., and Dean-Day, J.: Mapping hydroxyl variability throughout the global remote troposphere via synthesis of airborne and satellite formaldehyde observations, Proceedings of the National Academy of Sciences, 116, 11 171–11 180, https://doi.org/10.1073/pnas.1821661116, 2019.
  38. [39]Zhao, Y., Saunois, M., Bousquet, P., Lin, X., Berchet, A., Hegglin, M. I., Canadell, J. G., Jackson, R. B., Hauglustaine, D. A., Szopa, S., Stavert, A. R., Abraham, N. L., Archibald, A. T., Bekki, S., Deushi, M., Jöckel, P., Josse, B., Kinnison, D., Kirner, O., Marécal, V., O’Connor, F. M., Plummer, D. A., Revell, L. E., Rozanov, E., Stenke, A., Strode, S., Tilmes, S., Dlugokencky, E. J., and Zheng, B.: Inter-model comparison of global hydroxyl radical (OH) distributions and their impact on atmospheric methane over the 2000–2016 period, Atmospheric Chemistry and Physics, 19, 13 701–13 723, https://doi.org/10.5194/acp-19-13701-2019, 2019.

Paper details

vlnc:W_2z7hen8s
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents